Chapter 10 on today's products
- Approved by
- Panaversity ·
This page backs two lessons of Chapter 10: 10.3, on the route data takes, and 10.9, on carrying out one policy on today's tools. The lessons teach the differences that change a decision. This page holds each product's own settings, with their sources and the dates they were checked.
Products change every few weeks. Check the date on each box before you rely on a line, and open its source if the date is old.
10.3 The route matters as much as the product
The two boxes below show what each AI vendor's own pages said about training, temporary chats, retention, memory, records, connectors and usage policies.
Anthropic, as verified 18 November 2026.
- Training. On Free, Pro and Max, each user's privacy setting decides whether chats may be used to improve Claude. Feedback a user sends, and chats flagged for a safety review, can be used even with it off.1 With it on, chats can be kept in de-identified form for up to five years.2 On Team, Enterprise and the API, inputs and outputs are not used for training by default, unless a user sends feedback or the organization opts in.3
- Incognito chats. Not used for training, not saved to history, not used for memory, and kept for 30 days by default. Not available inside projects.4 On Team and Enterprise they are included in the organization's data exports, and on Enterprise in the Compliance API.4
- Retention. On individual plans, deleted chats are removed from back-end systems within 30 days, with exceptions. Chats flagged under the Usage Policy can be kept up to two years, and feedback for five years.2 Enterprise Owners can set a custom retention period of at least 30 days. Without one, Enterprise data is kept with no time limit.5
- Memory. On Team and Enterprise, Owners control whether memory is available. In most organizations created on or after October 31, 2026, it starts on for each member. In older ones it is off until the member turns it on.6
- Records. Audit logs are for Enterprise only, also through the Compliance API.7 On Team and Enterprise, the Primary Owner can export the organization's data.8
- Connectors. On Team and Enterprise, an Owner must enable a connector before members can use it. Owners can set each of a connector's tools to always allow, needs approval or blocked.9 Anthropic says custom connectors are not verified by Anthropic, so connect only to servers from organizations you trust. A connected service processes your data on its own systems, under its own terms, which may be outside the United States.9
- Usage Policy. The policy in effect since November 12, 2026 names eleven high-risk areas. They are legal, medical, finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits and services, and legal status and adjudication. Finance means a personal recommendation on investment products, advice on a person's own taxes, or preparing their tax return. A qualified person must review a recommendation in these areas before it is given as advice or used for a decision. That person must have the authority to change it, and remains responsible for what is delivered or decided. Anyone who receives the advice, or is the subject of the decision, must be clearly told that AI was used. General information, internal drafts, and business work that decides nothing about a specific person are outside these rules. Where the law allows, a few decisions wholly in the person's favor need no review first, such as paying an insurance claim in full.10
OpenAI, as verified 21 November 2026.
- Training. On plans for individuals, the "Improve the model for everyone" setting decides it. When it is off, OpenAI does not use new chats for training. Trusted Access sessions, such as Daybreak Blue, are an exception. On Business, Enterprise and Edu, business data is not used for training by default.11
- Temporary Chat. Not in history, does not create memories, not used to improve models. It may be kept for up to 30 days for safety.12 A chat cannot be added to a project while it is temporary.13
- Retention. Workspace admins on Business and Enterprise control how long data is kept.14
- Memory. Business workspace owners can turn memory and personalization off for the whole workspace.15 On Enterprise, admins control memory through workspace and role settings.16
- Records. Enterprise offers audit logs through the Compliance API, and an Admin Console with single sign-on and user provisioning.17
- Apps. On Business, apps are on by default and admins can change that. On Enterprise and Edu, new apps are off by default, with role-based access.18
- Usage Policies. They prohibit automating high-stakes actions in sensitive areas, including credit, employment, housing, legal, medical, insurance and education, without human review.19
10.9 The same policy on today's tools
The boxes above, and 10.7, gave each AI vendor's settings for training, retention, memory, connectors and apps. The two AI vendors' boxes below add the rest. They use the same six lines, in the same order, and a third box covers open source. Anthropic's adds one, for the one conversation.
Anthropic, as verified 18 November 2026.
- Usage Policy. The version in effect since November 12, 2026 is the one 10.3 summarizes. Its high-risk rules cover recommendations about one person, such as a credit, housing or employment decision. A qualified person must review such a recommendation, with authority to change it, before it is delivered or used. The person affected must be clearly told that AI was used. Internal drafts that are not the final recommendation are excluded. So is business work that advises or decides nothing about a specific person. Users must make sure their agents' actions, through tools, browsers or connected systems, follow the policy.10
- Your policy in the product. Owners on Team and Enterprise can write organization instructions, which Claude follows in every conversation. Only Owners and the Primary Owner can see them. Anthropic warns that in rare cases, when instructions directly contradict each other, behavior may vary.20
- Work accounts. Team and Enterprise offer single sign-on, so staff sign in with their work account.21 Once the company's email domain is verified, an Owner on either plan can stop people creating new personal accounts with it.21 On Enterprise, domain capture goes further. Personal accounts on the domain must move into the company's organization or close within 30 days, and the setting cannot be reversed.22
- Who switches features on. Of the standard roles, only Owners and the Primary Owner can turn on capabilities and integrations.23 Web search is one such capability.24 On Enterprise, custom roles give features to groups. But if an Owner turns a feature off for the whole organization, no custom role can turn it back on.25 Some settings start on. Members may use "Automatically approve", or Auto for short,26 in which Claude checks its own actions instead of asking, unless an Owner turns it off.24,27 A second setting, Start new tasks in "Automatically approve" mode, decides whether new tasks begin in it. It is on unless an Owner turns it off. Auto is rolling out as the default on Team. On Enterprise, new tasks still begin on "Manually approve", except in the Cowork side panel in Chrome.24 Owners can also turn off members' thumbs-up and thumbs-down ratings, the feedback that 10.3 says can be used for training.28
- The one conversation. On Team and Enterprise, it is available where the setting "Chat and Cowork unified" is on.29 On Enterprise, an Owner finds that setting under Organization settings, Capabilities. It is off by default in existing organizations. Anthropic's page says it turns on by default on November 16, 2026, unless an Owner has already chosen. On December 1, 2026, it also turns on for custom roles that have chat and code execution. That holds even if an admin turned it off for the role, where the organization is still on the default. It needs chat and cloud code execution to be on. It is not yet available with zero data retention, the HIPAA-ready configuration or customer-managed encryption keys. Custom roles can give it to chosen members, and Organization settings can also turn off "Skip all approvals".30 On Team, it turns on by default on December 1, 2026, unless an Owner keeps it off. New Team organizations already start with it on.31
- Records. On Enterprise, Owners can export the audit log for the past 180 days. It records events such as file uploads and deleted conversations, but not chat titles or content.7 Only the Primary Owner can turn on the Compliance API, which can also pull chat data and files.32
- Security documents. The Trust Center lists Claude Team and Enterprise as covered by SOC 2 Type 2 and ISO 27001. Sensitive documents need a request form.33
OpenAI, as verified 21 November 2026.
- Usage Policies. The current version took effect on November 20, 2026. It is the version that 10.3 summarizes. OpenAI says its rules are no substitute for legal requirements or professional duties. A form lets you follow updates.19
- Your policy in the product. On Enterprise, a workspace can set its own AI policy. ChatGPT shows it to users in a pop-up every 30 days, and again when the policy changes.34
- Work accounts. Business and Enterprise support single sign-on.35 Verifying the company's domain does not delete or merge anyone's personal workspace. Claiming the domain is only for approved use cases, by request to OpenAI.36 On Business, admins cannot require anyone to merge or delete a personal workspace.35 On Enterprise, joining can require one on the company's domain to be merged or deleted.37
- Who switches features on. On Enterprise, only Owners create and assign custom roles, which decide who may use each feature. Roles add up. If any role a member holds grants a feature, the member has it, even when another role turns it off. Lockdown Mode is the exception. OpenAI checks it separately, and it can block a feature that a role grants.38 If a business opts in to share data, for example through feedback, OpenAI may use it for training.14
- Records. The Compliance Platform is for Enterprise and Edu. Only a workspace Owner can grant access to conversation messages. Its logs are kept for 30 days, so a company that needs them longer should download and keep them itself.39
- Security documents. The Trust Portal offers its SOC 2 Type 2 report, for July 1, 2025 to June 30, 2026, to users who request access.40
Open source, as verified 23 October 2026.
- Where the conversation goes. An open-source tool can run on your own computer, and its records stay there. But it still sends the conversation to the company whose model it uses.41
- Data that must never leave. For that data, the only route is a model that runs on your own computer. OpenCode, an open-source coding agent, can use one through Ollama, LM Studio or vLLM.41
- Who switches features on. Each OpenCode agent starts from a rule that allows every action. By default, it asks only before reading sensitive environment files. Your own rules can set an action, such as editing files or running commands, to allow, ask or deny.42
- Running it safely. Running the tool yourself moves the security work to you. On a regular install, the OpenClaw assistant accepts connections only from its own computer.43 Its docs say to avoid opening it directly to the internet.44 Omnara, a runtime you can host yourself, has local settings that are insecure on purpose, so a real deployment follows its self-hosting guide.45
Keeping data in one region, as verified 18 November 2026 for Anthropic and 21 November 2026 for OpenAI. A rule that data must stay in one region is not a rule that it must never leave. With a region setting, the data still leaves your own systems. The setting decides where it is stored, or where the model works on it.
- Anthropic. On Anthropic's API, the
inference_geosetting can keep inference, the model's work on a request, in the US only. There is no EU option.46 Through Amazon Bedrock, a request can stay in one region you name, in Europe too.47 Through Google Cloud, one named region works only for Claude Sonnet 4.6 and earlier models. For newer models, a request can be kept within the United States or within the European Union.48 Microsoft Foundry offers a zone that keeps inference in the US.49 - OpenAI. New ChatGPT Enterprise and Edu customers can choose data residency, which stores their content in one of ten regions. Inference residency, in three of them, keeps the model's work there too. Some processing, such as signing in, can still happen outside the region. Features still in beta are not covered.50
- Outside services. A region setting does not reach them. OpenAI's page says an outside app, an MCP server or a web browsing provider handles what it receives under its own terms.50 So does a monitoring or tracing tool that a worker sends its steps to, on any product. Anthropic's page says such a tool stores what it receives under its own settings for access and retention.51
- Hosted runtimes. Data and traces, in Chapter 4 on today's products, shows which hosted runtimes keep a copy of your data, product by product. It also shows what a sandbox on your own infrastructure still sends to the vendor.
Sharing a result, as verified 18 November 2026 for Anthropic and 21 November 2026 for OpenAI. Section 10.10 asks who can open what the worker made. Each AI vendor's own pages answer part of it.
- Anthropic. Artifacts start private to you. On Team and Enterprise, they stay inside the organization by default. A member can share one with invited people or with everyone in the organization, as Can view, Commenter or Can edit. "Anyone with the link" is offered only when an Owner turns on external sharing, or allows that one artifact. Accepted invitations do not expire.52 In a shared project, members can see its knowledge and instructions. Each member's chats in it stay private unless that member shares them.53 A shared chat is a snapshot: later messages stay private unless you update it. On Team and Enterprise, only signed-in members of the same organization can see it.54
- OpenAI. A shared conversation link made in a Business, Enterprise or Edu workspace opens only for members of that workspace. Unlike a link from a personal account, it can include messages added after sharing. A link cannot be limited to chosen people and has no end date, so delete it when it is no longer needed. Anyone who can open a link can forward it. On Enterprise and Edu, admins can turn sharing off. Business has no such control for the whole workspace.55 In a shared project, members with chat access can see the project's chats, files and instructions.13
Check yourself
Question 1 / 2
0 answered
On ChatGPT Enterprise, one of a member's roles turns web search on and another turns it off. What happens?
Sources
- Is my data used for model training?, Anthropic Privacy Center.↑
- How long do you store my data?, Anthropic Privacy Center.↑↑2
- Is my data used for model training? Commercial products, Anthropic Privacy Center.↑
- Use incognito chats, Claude Help Center.↑↑2
- Configure custom data retention controls for Enterprise plans, Anthropic Privacy Center.↑
- Use Claude's chat search and memory to build on previous context, Claude Help Center.↑
- Access audit logs, Claude Help Center.↑↑2
- Export your organization's data, Claude Help Center.↑
- Use connectors to extend Claude's capabilities, Claude Help Center.↑↑2
- Usage Policy, Anthropic.↑↑2
- Data controls in ChatGPT, OpenAI Help Center.↑
- Temporary chat in ChatGPT, OpenAI Help Center.↑
- Projects in ChatGPT, OpenAI Help Center.↑↑2
- Enterprise privacy at OpenAI, OpenAI.↑↑2
- Memory FAQ (Business Version), OpenAI Help Center.↑
- Memory in ChatGPT, OpenAI Help Center.↑
- ChatGPT Enterprise admin quickstart, OpenAI Help Center.↑
- Admin controls, security, and compliance for plugins and apps, OpenAI Help Center.↑
- Usage policies, OpenAI.↑↑2
- Set organization instructions, Claude Help Center.↑
- Set up single sign-on (SSO), Claude Help Center.↑↑2
- Claim and migrate accounts on your domain, Claude Help Center.↑
- Roles and permissions, Claude Help Center.↑
- Use Claude Cowork on Team and Enterprise plans, Claude Help Center.↑↑2↑3
- Manage custom roles on Enterprise plans, Claude Help Center.↑
- Claude in Chrome permissions guide, Claude Help Center.↑
- Get started with Claude Cowork, Claude Help Center.↑
- Manage user feedback settings on Team and Enterprise plans, Claude Help Center.↑
- Claude Cowork and chat are one Claude, Claude Help Center.↑
- Set up and manage the new unified Claude experience for your Enterprise organization, Claude Help Center.↑
- Set up and manage the new unified Claude experience for your Team organization, Claude Help Center.↑
- Access the Compliance API, Claude Help Center.↑
- Trust Center, Anthropic.↑
- Managing workspace settings in ChatGPT Enterprise, OpenAI Help Center.↑
- Getting started with OpenAI identity and access for managed workspaces, OpenAI Help Center.↑↑2
- Verifying your domain for OpenAI identity, OpenAI Help Center.↑
- Managing members, workspace roles, and seats in ChatGPT Enterprise and Edu, OpenAI Help Center.↑
- Managing feature access with role-based access control in ChatGPT, OpenAI Help Center.↑
- OpenAI Compliance Platform for Enterprise and Edu customers, OpenAI Help Center.↑
- OpenAI Trust Portal, OpenAI.↑
- Providers, OpenCode v2 documentation.↑↑2
- Permissions, OpenCode v2 documentation.↑
- Security, OpenClaw documentation.↑
- Gateway exposure runbook, OpenClaw documentation.↑
- Omnara, README on GitHub.↑
- Data residency, Claude Platform Docs.↑
- Claude in Amazon Bedrock, Claude Platform Docs.↑
- Claude Models on Google Cloud, Claude Platform Docs.↑
- Claude in Microsoft Foundry, Claude Platform Docs.↑
- Data residency and inference residency for ChatGPT, OpenAI Help Center.↑↑2
- OpenTelemetry tracing for the Claude API, Claude Platform Docs.↑
- Manage project visibility and sharing, Claude Help Center.↑
- Public links for shared chats, Claude Help Center.↑
- Sharing conversations and scheduled tasks in ChatGPT, OpenAI Help Center.↑
- Owner
- Panaversity